#7171st of 90 in Open modelsFor image, video and design, video and animation
magi-1
Autoregressive video generation model with open weights and inference code
Live demo ↗
(opens in a new tab)Documentation ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 3.8k
- License
- Apache-2.0
- Last commit
- Jun 2026
- Language
- Python
3,654 on 25 Apr 20263,805 on 14 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
MAGI-1 generates video chunk by chunk (24 frames per chunk), with up to four chunks denoised concurrently. The repo has inference code and weights for 24B and 4.5B models, plus distilled and fp8-quantized variants. It supports text-to-video, image-to-video and video-to-video, driven by run.sh scripts and JSON configs.
Who it is for: Engineers with NVIDIA GPUs running or studying open video generation models
Strengths
- 4.5B models run on one RTX 4090; distill+fp8 needs 12GB with window_size 1
- Supports t2v, i2v and v2v, so existing video can be continued
- Distilled and fp8-quantized checkpoints lower hardware needs for the 24B model
- Apache-2.0 code, with base, distill and quant weights on Hugging Face
Weaknesses
- 24B models need 8x H100/H800; quantized needs 4x H100 or 8x RTX 4090
- No web UI or API server; inference runs through run.sh scripts
- Source install pins Python 3.10, PyTorch 2.4 and CUDA 12.4
- No tagged releases; the Docker command uses --privileged and host networking
What it needs
- GPU required
- Needs PyTorch 2.4.0, CUDA 12.4, ffmpeg 4.4, MagiAttention (optional, Hopper GPUs), T5 and VAE weights from Hugging Face
- Models: MAGI-1 24B, MAGI-1 4.5B, MAGI-1.1 24B, distill, distill+fp8_quant
Also in Open models
See all 90| Rank | Project | Score |
|---|---|---|
| 1 | ruviewWiFi CSI sensing on ESP32 for presence, breathing, heart rate and pose | 94 out of 100 |
| 2 | layaSingle-pass multilingual classifier returning typed decisions with calibrated odds | 88 out of 100 |
| 3 | rf-detrReal-time transformer models for object detection, segmentation and keypoints | 82 out of 100 |
| #4 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 80 out of 100 |
| #5 | open_clipPyTorch library for training and running CLIP-style image-text models | 79 out of 100 |